Download presentation

Presentation is loading. Please wait.

Published byGarrison O'Neal Modified over 3 years ago

1
Quantifying and managing uncertainty with Gaussian process emulators Tony O’Hagan University of Sheffield

2
Simulators 7/12/2011mucm.ac.uk NAFEMSSlide 2 In almost all fields of science, technology, industry and policy making, people use mechanistic models to simulate complex real-world processes For understanding, prediction, control Usually implemented in computer codes Often very computationally intensive We’ll call them simulators There is a growing realisation of the importance of uncertainty in simulator predictions Can we trust them? Without any quantification of output uncertainty, it’s easy to dismiss them

3
Examples Slide 3 Climate prediction Molecular dynamics Nuclear waste disposal Oil fields Engineering design Hydrology 7/12/2011mucm.ac.uk NAFEMS

4
Sources of uncertainty 7/12/2011mucm.ac.uk NAFEMSSlide 4 A simulator takes inputs x and produces outputs y = f(x) How might y differ from the true real-world value z that the simulator is supposed to predict? Error in inputs x Initial values, forcing inputs, model parameters Error in model structure or solution Wrong, inaccurate or incomplete science Bugs, solution errors

5
Quantifying uncertainty 7/12/2011mucm.ac.uk NAFEMSSlide 5 The ideal is to provide a probability distribution p(z) for the true real-world value The centre of the distribution is a best estimate Its spread shows how much uncertainty about z is induced by uncertainties on the last slide How do we get this? Input uncertainty: characterise p(x), propagate through to p(y) For example, use Monte Carlo sampling Generate random sample of x values from p(x), run the model for each to get a random sample from p(y) Structural uncertainty: characterise p(z-y)

6
Reducing uncertainty To reduce uncertainty, get more information! Informal – more/better science Tighten p(x) through improved understanding Tighten p(z-y) through improved modelling or programming Formal – using real-world data Calibration – learn about model parameters Data assimilation – learn about the state variables Learn about structural error z-y Validation mucm.ac.uk NAFEMS7/12/2011

7
So far, so good 7/12/2011mucm.ac.uk NAFEMSSlide 7 In principle, all this is straightforward In practice, there are many technical difficulties Formulating uncertainty on inputs Elicitation of expert judgements Propagating input uncertainty Modelling structural error Anything involving observational data! The last two are intricately linked And computation

8
The problem of big models Slide 8 Key tasks require us to run the simulator many times Uncertainty propagation Implicitly, we need to run f(x) at all possible x Monte Carlo works by taking a sample of x from p(x) Typically needs thousands of simulator runs Calibration Learn about uncertain inputs from observations of the real process Traditionally this is done by searching the x space for good fits to the data These techniques are impractical if the simulator takes more than a few seconds to run We need a more efficient technique 7/12/2011mucm.ac.uk NAFEMS

9
Gaussian process representation 7/12/2011mucm.ac.uk NAFEMSSlide 9 More efficient approach First work in early 1980s Consider the simulator as an unknown function f(.) becomes a random process We represent it as a Gaussian process (GP) Conditional on hyperparameters Or its Bayes linear analogue Training runs Run simulator for sample of x values Condition GP on observed data Typically requires many fewer runs than MC One to three orders of magnitude fewer And x values don’t need to be chosen randomly

10
Emulation 7/12/2011mucm.ac.uk NAFEMSSlide 10 Analysis is completed by prior distributions for, and posterior estimation of, hyperparameters The posterior distribution is known as an emulator of the simulator Posterior mean estimates what the simulator would produce for any untried x (prediction) With uncertainty about that prediction given by posterior variance Correctly reproduces training data

11
2 code runs 7/12/2011mucm.ac.uk NAFEMSSlide 11 Consider one input and one output Emulator estimate interpolates data Emulator uncertainty grows between data points

12
3 code runs 7/12/2011mucm.ac.uk NAFEMSSlide 12 Adding another point changes estimate and reduces uncertainty

13
5 code runs 7/12/2011mucm.ac.uk NAFEMSSlide 13 And so on

14
Then what? 7/12/2011mucm.ac.uk NAFEMSSlide 14 Given enough training data points we can emulate any simulator accurately So that posterior variance is small “everywhere” Typically, this can be done with orders of magnitude fewer simulator runs than traditional methods Use the emulator to make inference about other things of interest Uncertainty analysis (Wright, Brown) Sensitivity analysis (Vogt, Thole, Doyle), Calibration, data assimilation, validation Optimisation Whatever you would like to do with the simulator And emulator results have quantified uncertainty

15
Example: UK carbon flux in 2000 7/12/2011mucm.ac.uk NAFEMSSlide 15 Vegetation simulator predicts carbon exchange from each of 700 pixels over England & Wales Principal output is Net Biosphere Production Sheffield Dynamic Global Vegetation Model (SDGVM) Accounting for uncertainty in inputs Soil properties Properties of different types of vegetation Propagate input uncertainty through the model Aggregated to England & Wales total Allowing for correlations Estimate 7.61 Mt C Std deviation 0.61 Mt C

16
Maps 7/12/2011mucm.ac.uk NAFEMSSlide 16

17
England & Wales aggregate 7/12/2011mucm.ac.uk NAFEMSSlide 17 PFT Plug-in estimate (Mt C) Mean (Mt C) Variance (Mt C 2 ) Grass5.284.650.323 Crop0.850.500.038 Deciduous2.131.690.009 Evergreen0.800.780.001 Covariances0.001 Total9.067.610.372

18
Role of emulation 7/12/2011mucm.ac.uk NAFEMS18 Gaussian process emulation was crucial to the feasibility of this exercise Almost 3000 simulator runs for a single set of inputs Imagine this repeated hundreds or thousands of times for Monte Carlo And all that repeated to evaluate the sensitivity to each input group We emulated each PFT at a sample of 33 sites Typically 200 simulator runs for each Kriging to interpolate between sites Also equivalent to Gaussian process emulation Kennedy, M. C. et al (2008). Quantifying uncertainty in the biospheric carbon flux for England and Wales. Journal of the Royal Statistical Society A 171, 109-135.

19
Alternative methods 7/12/2011mucm.ac.uk NAFEMS19 Monte Carlo And refinements like LHC sampling Inefficient Alternative surrogates Response surfaces, neural nets, etc. All approximate and simplify Whereas emulator encapsulates knowledge exactly Internal error measures are wrong UQ methods Polynomial chaos, stochastic collocation etc. (Powell) ? Lack usable internal error measures ? Limited range of tasks

20
Resources 7/12/2011mucm.ac.uk NAFEMS20 MUCM project Managing Uncertainty in Complex Models http://mucm.ac.uk http://mucm.ac.uk Advisory Panel provides industrial/community involvement MUCM toolkit Large set of web pages on building and using emulators Background, theory, discussion, advice, procedures, examples Case studies UCM community mailing list http://mucm.ac.uk/Pages/UCM.html http://mucm.ac.uk/Pages/UCM.html UCM 2012 conference (2-4 July in Sheffield) http://mucm.ac.uk/UCM2012.html http://mucm.ac.uk/UCM2012.html

Similar presentations

Presentation is loading. Please wait....

OK

Kriging.

Kriging.

© 2018 SlidePlayer.com Inc.

All rights reserved.

To make this website work, we log user data and share it with processors. To use this website, you must agree to our Privacy Policy, including cookie policy.

Ads by Google

Ppt on atrial septal defect in adults Convert doc file to ppt online maker Ppt on seed production technology Ppt on solar power plant in india Ppt on unipolar nrz Ppt on tourism in delhi and its promotion Adrenal gland anatomy and physiology ppt on cells Ppt on waxes are lipids Person focused pay ppt online Ppt on edge detection in c